perf: add ASCII-safe substr fast path

[He-Pin]

Motivation:
Long printable ASCII strings can use UTF-16 length and substring indexes directly because every UTF-16 code unit is also one Jsonnet codepoint. std.substr currently still pays codepoint offset checks for these proven-safe long literals.

Key Design Decision:
Reuse the existing _asciiSafe runtime marker, but only set it for long (>1024 chars) string literals that are printable ASCII and JSON-render safe. This keeps parser overhead off short-string-heavy workloads and avoids broad runtime ASCII scans. JVM uses byte-level SWAR for the long-string safety scan; Scala Native and Scala.js keep a scalar single-pass scan because the byte-buffer SWAR variant regressed Native due UTF-8 allocation cost.

Modification:

• Add CharSWAR.isAsciiJsonSafe on JVM, Native, and JS.
• Use JVM byte-SWAR to check long strings for ASCII + JSON-safe bytes.
• Mark long ASCII JSON-safe literals in Parser.makeStr.
• Propagate _asciiSafe through Val.Str.concat when both operands are safe.
• Let std.length and std.substr use direct String.length / substring only for proven-safe strings.
• Add Unicode/ASCII regression coverage for long ASCII length, substr bounds, and concat propagation.

Benchmark Results:
JMH, lower is better (./mill --no-server -j 1 bench.runRegressions ...):

benchmark	master ms/op	this PR ms/op	result
go_suite/substr.jsonnet	0.056	0.045	~20% faster
jdk17_suite/split_resolve.jsonnet	0.147	0.146	neutral
cpp_suite/realistic2.jsonnet	44.529	45.024	neutral/no stable regression signal

Scala Native hyperfine, lower is better (--shell=none --warmup 5 --runs 20):

command	mean	note
master native go_suite/substr.jsonnet	7.7 ± 1.8 ms	baseline
this PR native go_suite/substr.jsonnet	6.5 ± 1.0 ms	neutral/slightly faster vs master
jrsonnet go_suite/substr.jsonnet	4.0 ± 0.6 ms	still ahead

Analysis:
The broad #776 ASCII/string fast-path ideas were narrowed aggressively. Runtime ASCII scans in std.length/std.substr caused guard regressions, and all-platform UTF-8 byte-SWAR regressed Scala Native. This version uses byte-SWAR only where it measured well (JVM), keeps Native/JS allocation-free, and only takes fast paths when the marker is already proven, preserving Jsonnet codepoint semantics for all non-ASCII strings.

References:

• Source idea: perf: comprehensive Scala Native render pipeline optimization #776
• Narrowed from commit: He-Pin@a190a800

Result:
Draft PR for review. JVM substr benchmark is positive; Native is neutral/slightly positive with no-shell hyperfine; full ./mill --no-server -j 1 __.reformat && ./mill --no-server -j 1 __.test passed locally.

[stephenamar-db]

rebase

[He-Pin]

@stephenamar-db rebased